| commit | ef481a44f57a8ad83b79b7a7ca6319c8d7123e2a | |
| author | Mark Abraham <mark.j.abraham@gmail.com> | |
| Mon, 21 Aug 2017 16:11:05 +0000 (21 18:11 +0200) | ||
| committer | Berk Hess <hess@kth.se> | |
| Wed, 23 Aug 2017 05:55:19 +0000 (23 07:55 +0200) | ||
| tree | b34a90708fd9c018713c059989b07121d9782eb5 | treesnapshot (tar.gz zip) |
| parent | 795ad1f89751d8d7e8832a28fcfdf9e53d85d0db | commitdiff |
Stopped checkpoint influencing parallelism setup
There's no strong reason to prefer that by default the parallelism
setup is the same as it was for the run in the checkpoint. If the next
run has the same number of ranks, then either we should follow the
user's instructions for the second run, or apply the defaults. This
permits us to fix bugs in the default settings for patch releases, and
have it work. It also prepares for future auto-tuning when we hope to
be able to adapt to the run-time environment, which will become
increasingly volatile, between and within runs. And of course it
simplifies the setup code slightly.
Change-Id: I41d65132f0855cd57a0a572d4579045bf10f295e
There's no strong reason to prefer that by default the parallelism
setup is the same as it was for the run in the checkpoint. If the next
run has the same number of ranks, then either we should follow the
user's instructions for the second run, or apply the defaults. This
permits us to fix bugs in the default settings for patch releases, and
have it work. It also prepares for future auto-tuning when we hope to
be able to adapt to the run-time environment, which will become
increasingly volatile, between and within runs. And of course it
simplifies the setup code slightly.
Change-Id: I41d65132f0855cd57a0a572d4579045bf10f295e
| src/gromacs/fileio/checkpoint.cpp | diffblobblamehistory | |
| src/gromacs/fileio/checkpoint.h | diffblobblamehistory | |
| src/programs/mdrun/runner.cpp | diffblobblamehistory |